Practical Sandboxing on the Windows Platform An assessment of the Internet Explorer, Adobe Reader and Google Chrome sand

1. Practical Sandboxing on the Windows Platform An assessment of the Internet Explorer, Adobe Reader and Google Chrome sandboxes By Tom Keetch

2. About Me • Verizon Business • Lead consultant for Code Review in EMEA • Previous Presentations • OWASP AppSec EU 2011 – Practical Sandboxing with Chromium • CONfidence 2011 - Assessing Practical Sandboxes (Updated) • BlackHat Europe 2011 – Assessing Practical Sandboxes • Hack.LU 2010 - Protected Mode Internet Explorer • Exploit mitigations are my favourite topic! • How to make exploits prohibitively expensive to find and exploit…

3. Introduction • What is Practical Sandboxing? • User-mode sandboxing methodology • Based on Windows OS facilities • Overview of 3 implementations: • Protected Mode Internet Explorer (limited) • Adobe Reader X • Chromium • This presentation is about: • Breaking out of such Sandboxes with the minimum required effort.

4. Agenda • Sandboxes for exploit mitigation (Theory) • Overview of Practical Sandboxing Implementations (Background) • Sandboxing Flaws (Practical) • A counter-argument to Adobe’s view of their sandbox as an exploit mitigation (Argumentative) • Conclusions

5. Sandboxes for Exploit Mitigation

6. Sandboxes for exploit mitigation • Two options for exploit mitigation: • Increase cost of exploitation (DEP, ASLR) • Decrease target value (Sandboxes) • But a second stage exploit, can usually bypass the sandbox for finite cost... • This presentation focuses on sandbox-escape. • Please read the whitepapers for more information. • Too much to cover in just 1 hour • Come talk to me after!

7. “Return-on-Exploitation”

8. Two Potential Failures • The cost of bypassing the exploit mitigation is too low to deter a potential attacker. • Trivial to bypass? • High Target Value? • The reduction of value of the target is not sufficient to deter a potential attacker. • Protecting the wrong assets? • Some assets cannot be protected by a sandbox.

10. Looking for “cheap” exploits • This research set out to find the easiest places to find sandbox-escape exploits. • Cheap-to-find exploit types were found: • Previously unexposed interfaces • Easily detectable (and exploitable) conditions • Mostly due to incomplete sandbox implementations… • … but also sandbox-specific issues. • Also, resources not protected by sandbox: • Network Access • Resources protected by the Same Origin Policy • Kernel Interfaces

11. Overview of Practical Sandbox Implementations

12. The Practical SandboxingMethodology • Restricted Access token • Deny-only SIDs (Discretionary) • Low Integrity (Mandatory) • Privilege Stripping (Capability) • GUI Sub-System Security • Window Station Isolation • Desktop Isolation • Job Object Restrictions Each mitigates against a different attack!

13. Protected Mode Internet Explorer • Session • WinStation • Desktop • Medium Integrity Internet Explorer (Broker) • Low Integrity Browser Tab (Internet Zone) Browser Tab (Trusted Zone, Local Intranet Zone)

14. Protected Mode Internet ExplorerPractical Sandboxing Check-list

15. Protected Mode Internet ExplorerSandboxing • Sandbox Limitations: • Only supported on Vista and later, because only Integrity Levels are used. • Only protected the Integrity of the system, not confidentiality. • Full access to Windows station resources (including Clipboard, GAT). • Many possible sandbox escape routes including: • UAC Launches • Trusted Broker attacks • Generic PMIE bypass for a domain-joined workstation. • More information previously presented at Hack.LU, Oct 2010. • Not a Security Boundary, for many reasons. • Lots of potential elevation routes.

16. Adobe Reader X • Session • Medium Integrity • (Workstation) • (Desktop) Adobe Reader (Broker) PDF Renderer • Restricted Token • Low Integrity • Job Object

17. Adobe Reader XPractical Sandboxing Check-list

18. Adobe Reader X Sandboxing • Makes use of Chromium sandboxing and IPC framework (BSD license) • PDF Rendering is sandboxed. • Sandbox Limitations: • The broker does not restrict read access. • Sandbox doesn't protect user’s clipboard • Full Access is granted to the Global Atom Table. • “NT Authority\RESTRICTED” SID -> BNO Namespace squatting. • No WinSta or Desktop isolation, but compensated for with Job Object restrictions. • Read Adobe Blog posts for more information.

19. Chromium • Session • Medium Integrity Google Chrome (Broker) GPU Process Browser Tab Plug-in • Restricted Token • Low Integrity • Job Object • Per plug-in sandboxing.

20. ChromiumPractical Sandboxing Check-list *Renderer sandbox.

21. Chromium sandboxing • A flexible framework for applying the full “practical sandboxing” methodology • Renderer is in the most restrictive possible sandbox. • 3rd Party Plug-ins are often not sandboxed • Shockwave, Java etc. • Adobe Flash has a *limited* sandbox. • GPU process is not sandboxed (planned for future release)

22. The VUPEN Exploit (I)

23. The VUPEN Exploit (II)

24. The VUPEN Exploit (III)

25. The Adobe Flash Sandbox Bypass using “shatter attacks”!

26. Cheap Exploit Vectors

27. Cheap Exploit Vector #1

28. BNO Namespace Squatting • Shared sections can be created with a name in the 'Local' namespace • Shared Sections • Mutexes, Events, Semaphores (Synchronisation objects) • By “squatting” on named object, we can set arbitrary permissions on the object if: • It can be created before the application • If the application does not fail if the named object already exists. • If we know or can predict the name of the object. • This can expose applications outside the sandbox to attacks they never knew existed…

29. BNO Namespace Squatting –PMIE Sandbox-Escape • Terminate the Medium IL iexplore.exe process. • Predict the PID of the new process. • Create the “ie_lcie_main_<pid>” shared section. • Initialise the section with malicious data. • When iexplore.exe initialises LCIE, malicious code will execute outside of the sandbox.

30. The Fuzzer that found it... int _tmain(int argc, _TCHAR* argv[]) { unsigned int size = _tstoi(argv[2]); HANDLE hSection = CreateFileMapping(NULL, NULL, PAGE_EXECUTE_READWRITE, 0, size, argv[1]); unsigned char* lpBuff = (unsigned char*) MapViewOfFile(hSection, FILE_MAP_WRITE | FILE_MAP_READ, 0, 0, size); // Take a copy of the initial contents of the section. memcpy(init, lpBuff, size); while(1) { memcpy(lpBuff, init, sizeof(init)); for(unsigned int i = 32; i < size; i++) if(rand() % 1000 < 5 ) lpBuff[i] = (unsigned char) rand(); PROCESS_INFORMATION ProcInfo1 = {0}; STARTUPINFOA StartupInfo1 = {0}; CreateProcessA(NULL, "C:\\Program Files\\Internet Explorer\\iexplore.exe", NULL, NULL, FALSE, 0, NULL, NULL,&StartupInfo1, &ProcInfo1); CloseHandle(ProcInfo1.hProcess); CloseHandle(ProcInfo1.hThread); Sleep(2000); PROCESS_INFORMATION ProcInfo2 = {0}; STARTUPINFOA StartupInfo2 = {0}; CreateProcessA(NULL, "pskill iexplore.exe", NULL, NULL, FALSE, 0, NULL, NULL, &StartupInfo2, &ProcInfo2); CloseHandle(ProcInfo2.hProcess); CloseHandle(ProcInfo2.hThread); Sleep(1000); } return 0; }

31. MSRC’s Response “As we are able to reproduce the crashes I have asked the Internet Explorer product team to address this issue in a next release of IE, which will most likely be IE10 rather than IE9 as that version is pretty much complete” … “We decided to close the case because Protected Mode IE is not presently a security boundary, thus a sandbox escaping is not considered a security vulnerability.”

32. BNO Namespace Squatting –Other Practical Sandboxes • Exploitable from Protected Mode IE • Low Integrity only • Exploitable from Adobe Reader X • Process Token has “NT AUTHORITY\RESTRICTED” • Grants access to create objects in Local BNO • Not Exploitable from Chrome Renderer sandbox • “Naked” Process Token. • But if Microsoft won’t fix this bug until IE 10? • How many more similar vulnerabilities are there? • Adobe Reader and other incomplete implementations will remain vulnerable to BNO Namespace Squatting issues.

33. Cheap Exploit #2

34. NPAPI Interface Exploits(Chromium Specific) • NPAPI was originally used to interface between the Netscape browser and an in-process plug-in. Process Browser Browser Tab NPAPI Plug-in (DLL) NPAPI

35. Out-of-Process NPAPI • Later NPAPI crossed process boundaries • Improved stability, no improved security. Session Google Chrome (Broker Process) Browser Tab (Process) Trusted Plug-in (Process) NPAPI

36. NPAPI In Chrome (Today) • NPAPI now crosses a security boundary between sandboxed tabs and un-sandboxed plug-ins. Session Google Chrome (Broker Process) Browser Tab (Process) Trusted Plug-in (Process) NPAPI Sandbox

37. NPAPI Exploits • NPAPI Callers were previously trusted... • ...Now they are not. • Flash and other plug-ins are currently not securely sandboxed. • Exploitable bugs in Adobe (and other vendors) code will allow sandbox-escape. • But these bugs were previously not vulnerabilities → Implicit Calling conventions?

38. A benign crash? • Thread 9 *CRASHED* ( EXCEPTION_ACCESS_VIOLATION @ 0x09ccf232 ) 0x102e5c06 [NPSWF32.dll - memcpy.asm:257] memcpy 0x102e1828 [NPSWF32.dll + 0x002e1828] CBitStream::Fill(unsigned char const*, int) 0x102e0b96 [NPSWF32.dll + 0x002e0b96] mp3decFill 0x102e0892 [NPSWF32.dll + 0x002e0892] PlatformMp3Decoder::Refill(int,unsigned char*) 0x10063d21 [NPSWF32.dll + 0x00063d21] CMp3Decomp::GetDecompressedData(short*,int,int,int,int) 0x10063f62 [NPSWF32.dll + 0x00063f62] CMp3Decomp::Decompress(short *,int) 0x100ad448 [NPSWF32.dll + 0x000ad448] CoreSoundMix::BuildBuffer(int) 0x100ae2c5 [NPSWF32.dll + 0x000ae2c5] CoreSoundMix::SendBuffer(int,int) 0x10153d6b [NPSWF32.dll + 0x00153d6b] PlatformSoundMix::SoundThread() 0x10154034 [NPSWF32.dll + 0x00154034] PlatformSoundMix::SoundThreadFunc(void *) 0x7c80b728 [kernel32.dll + 0x0000b728] BaseThreadStart Full report @ http://crash/reportdetail?reportid=b370c132fc6587f7 Google Chrome 4.0.249.70 (Official Build 36218) • This was found by accident (using Chromium) • Fixed by Adobe!

39. Input events • Can also send key and mouse events. • NPP_InputEvent(). • Possible to bypass Flash Security Dialogs • Enable web-cam • Enable Microphone • Plug-ins are currently unable to distinguish between user input and simulated input from renderer. • Hopefully this is fixed in “Pepper”.

40. Cheap Exploit #3

41. Handle Leaks • Handles which refer to privileged resources may exist in sandboxes for several reasons. • A handle can be used for any operation for which it has already been granted access. • If the right type of handle is leaked into the sandbox, it can be used for sandbox-escape. • These handles are easily detected at run-time!

42. What causes “Handle Leaks”? • Deliberately granted by broker. • Accidentally granted by broker. • Incorrectly granted by broker (policy error) • Unclosed handles from sandbox initialisation • Before Lock-down (init. with unrestricted token) • Internal handles kept open by libraries • Internal handles kept open by 3rd Party Hook DLLs • Libraries in use must be “sandbox friendly”

43. Adobe Reader X Handle Leaks • Sandboxed renderer has write access to the Medium-integrity Internet Explorer cookie store, history etc. • The ARX broker also doesn't currently restrict read access to local file system.

44. Cheap Exploit #4

45. Clipboard Attacks • In PMIE and AR-X, the clipboard is shared between the sandbox and the rest of the user's session. • Ever put your password in the clipboard? • What about attacking other applications? • Previously, the clipboard contents were normally trustworthy, now they are not.

46. Clipboard Attacks • What about... • Pasting malicious command lines into a shell followed by a new line? • Inputting maliciously formatted data into the clipboard? • Do application developers implicitly trust clipboard contents?

47. A counter-argument to Adobe’sview of the sandbox